# PrismML Compresses Qwen3.8-27B to 5.9 Gigabytes While Keeping 98.2% of Its Benchmark Average

Ternary weights cut the model to 1.76 effective bits, but long-horizon agent tasks degrade sharply, with Terminal-Bench 2.1 falling to 52.8 from 69.7.

- Published: 2026-09-20T06:20:15.045Z
- Canonical: https://polylog.news/ai/2026-09-20/prismml-compresses-qwen3-8-27b-to-5-9-gigabytes-while-keepin
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [MarkTechPost](https://www.marktechpost.com/2026/09/18/prismml-releases-ternary-bonsai-2-27b-a-5-9-gb-apache-2-0-model-retaining-98-2-of-qwen3-8-27b-performance/), [PrismML](https://prismml.com/news/bonsai-2-27b)

PrismML, a startup founded by a team out of Caltech, released Ternary Bonsai 2 27B, a compressed version of Alibaba's Qwen3.8-27B. Each weight is stored as one of three values, minus one, zero or plus one, with FP16 group-wise scaling, giving 1.76 effective bits per weight. The full-precision reference occupies 53.81 gigabytes. The compressed model occupies 5.93 gigabytes, [about nine times smaller](https://www.marktechpost.com/2026/09/18/prismml-releases-ternary-bonsai-2-27b-a-5-9-gb-apache-2-0-model-retaining-98-2-of-qwen3-8-27b-performance/), and the weights carry an Apache 2.0 license.

Across a 20-benchmark suite covering reasoning, math, coding, instruction following, vision and agentic tool use, PrismML reports an aggregate score of 83.9 against 85.4 for the original, or 98.2% retention. The Russian-language channel AI ML Big Data [carried the same figures](https://t.me/ai_machinelearning_big_data/10969).

The aggregate number hides where the loss concentrates. PrismML's own breakdown puts math and coding at parity, vision retention at 96.3%, and knowledge and reasoning at 96.9%. Long-horizon agent work falls much further. On Terminal-Bench 2.1 the compressed model scores 52.8 against 69.7, and on SWE-bench Verified it scores 60.8 against 80.6. Those are the two evaluations that most closely track what an agent framework actually asks a model to do over many turns, and a drop of roughly 20 points on each is not a rounding error.

All of these numbers come from the vendor. No independent reproduction has been published. The compression itself is verifiable by anyone who downloads the weights, which run on a 16-gigabyte laptop or a single 24-gigabyte graphics processor, so the claim is unusually easy to check.

## What this means

Aggregate retention scores are the wrong metric for buyers building agents. A model that holds parity on single-turn math and coding while losing 20 points on multi-turn terminal and repository tasks is suitable for local chat, classification and retrieval, and unsuitable for unattended agent loops, and a single headline percentage hides that split. The gainers are laptop and workstation silicon vendors and anyone serving latency-sensitive inference on-premises, since a 27-billion-parameter class model now fits in consumer memory. The losers are API vendors whose lowest tiers exist to serve exactly the summarize-classify-extract workloads that survive ternary compression intact.

## What to watch

- Whether independent evaluators reproduce the Terminal-Bench 2.1 and SWE-bench Verified gaps, since the size of that agentic penalty determines whether compressed models can be used unattended at all.
- Whether Alibaba or other open-weight publishers start shipping ternary variants themselves, which would move extreme quantization from a third-party service into the base release.
- Whether hosted providers list the compressed model at a price low enough to undercut small closed models, which is the channel through which this reaches most developers.